# Frozen contextual teacher: integrated development readout

Descriptive adjunct prepared while G was queued, before native generation. No additional gates or inferential claims. Combined gradients do not measure component dose or Adam displacement. Retention is historical source likelihood, not native capability. Exposure is treatment-dependent, not matched causal strata.

| Policy | Correct / 512 |
|---|---:|
| W | 238 |
| G | 261 |
| F | 236 |
| E | 236 |

Frozen development spending gate: **False**.

| Contrast | Net correct | Difference, pp | Paired 95% interval, pp |
|---|---:|---:|---|
| F-W | -2 | -0.39 | [-4.30, 3.52] |
| F-G | -25 | -4.88 | [-8.59, -1.17] |
| F-E | 0 | 0.00 | [-3.71, 3.71] |

| Arm | Optimizer updates / scheduled | Donor groups / questions | Receiver draws / draws | Dense tokens | Clipped batches / scheduled | Source NLL change |
|---|---:|---:|---:|---:|---:|---:|
| G | 70/80 | 182/320 | 326/2560 | 0 | 70/80 | +0.038809 |
| F | 54/64 | 151/256 | 237/2048 | 83984 | 54/64 | +0.024981 |
| E | 55/64 | 154/256 | 261/2048 | 94065 | 55/64 | +0.015480 |

| Arm | Training allocated GPU seconds | Matrix + attention proxy, FLOPs |
|---|---:|---:|
| G | 3312.2 | 3.3933e+15 |
| F | 2747.6 | 3.1442e+15 |
| E | 2741.4 | 3.0843e+15 |

| Policy | Evaluation GPU seconds | Generated tokens |
|---|---:|---:|
| W | 1029.7 | 305807 |
| G | 1004.9 | 308305 |
| F | 965.5 | 301013 |
| E | 1026.1 | 320488 |

Total training/evaluation allocation including interrupted work: [12863.60, 12863.60] GPU seconds. Any nonzero interval width reflects missing interruption exit timing, not statistical uncertainty. Readiness57.0s and freshWarm460.1s remain additional, as do older research ancestry costs.

Interrupted attempts:

- notes/stable_contextual_interruption_20261013_migration/interruption.json: 0 retained draws, 0 uncheckpointed updates, 0 retained generated tokens. This work was repeated and remains charged.

Completed jobs only; logical matrix/attention arithmetic proxy excludes norms, embeddings, nonlinearities, transfers, optimizer, grader and serialization. Checkpoint recomputation charged. Not measured hardware FLOPs.

Reused512 development; one seed/start; not compute matched; conditional paired item intervals exclude seed uncertainty; no benchmark or sustained-loop claim

Primary claims follow the unchanged frozen contrasts. F−E is the fixed contextual-versus-empty contrast. Frozen teacher identity does not make visited prefixes or coarsened target distributions fixed. No current/stable interaction or semantic-content-isolation claim follows. Shared ancestry and source costs are additional, as recorded in the JSON ledger.
